Jointly Learning Non-Cartesian k-Space Trajectories and Reconstruction Networks for 2D and 3D MR Imaging through...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Compressed sensing in MRI optimizes k-space sampling and image reconstruction from undersampled data.
• Deep learning methods can address both sampling and reconstruction simultaneously, particularly for non-Cartesian acquisitions.

Purpose of the Study:

• To develop a novel deep learning approach for k-space trajectory optimization in MRI.
• To improve hardware-compliant sampling and image reconstruction quality in accelerated MRI.

Main Methods:

• A projection-based deep learning method was developed to enforce MR hardware constraints for k-space trajectories.
• Ablation studies compared the projection method against penalty-term based approaches.
• The method was extended to 3D, comparing data-driven joint learning with model-based methods.

Main Results:

• The projection-based scheme demonstrated superior k-space trajectory flexibility and improved reconstructed image quality.
• In 2D retrospective studies on fastMRI, a 20-fold acceleration achieved SSIM scores of 0.92-0.95 and a 3-4 dB PSNR gain.
• In 3D, the data-driven joint learning method outperformed model-based methods like SPARKLING with a 2 dB PSNR and 0.02 SSIM gain.

Conclusions:

• A projection-based deep learning approach offers enhanced flexibility and performance for accelerated MRI acquisition.
• This method significantly improves image reconstruction quality compared to current state-of-the-art techniques.
• The approach is effective in both 2D and 3D MRI, showing promise for clinical applications.